Question Answering Systems are computational systems designed to provide answers to questions expressed in natural language or another structured form. Unlike conventional search systems, which generally return a ranked list of documents or passages, a Question Answering System attempts to identify or construct an answer that directly addresses the question. The answer may be extracted verbatim from a source, selected from a structured knowledge base, generated from retrieved evidence or produced through a combination of these methods.
Question answering can be open-domain, where questions may concern almost any subject, or closed-domain, where the system operates within a restricted knowledge source such as a medical database or technical manual. The task combines language understanding, information retrieval, knowledge representation and, in modern systems, natural language generation. A reliable system must not only interpret the question but also identify appropriate evidence and formulate an answer consistent with that evidence.
Early Question Answering Systems often operated over structured knowledge bases or tightly controlled domains. Systems such as database question-answering interfaces translated natural-language questions into structured queries, while reading-comprehension systems attempted to locate answers within documents. Open-domain question answering subsequently became closely connected with Information Retrieval: a system could retrieve potentially relevant documents and then search those documents for answer-bearing passages.
The development of large machine-readable knowledge bases enabled another approach in which questions were mapped to entities, relations or database queries. Statistical machine learning improved answer selection, while neural networks enabled more flexible representations of questions and passages. Transformer-based language models substantially expanded the ability of systems to interpret questions and produce answers, particularly when combined with retrieval mechanisms that provide external evidence.
Question Answering Systems can be divided into extractive, abstractive, generative and structured approaches, although practical systems may combine them. Extractive question answering selects an answer span from a supplied passage or document, making the source of the answer relatively straightforward to identify. Abstractive or generative systems compose an answer that may not appear verbatim in the source. Knowledge-base question answering translates a question into a query over structured facts, while hybrid systems can combine structured and unstructured evidence.
Retrieval-augmented question answering first retrieves relevant passages and then uses an answering model to interpret them. This architecture is useful when information changes over time because the retrieval corpus can be updated independently of the model's original training. The distinction between answering from retrieved evidence and answering from internal model knowledge is important because their failure modes and verification requirements differ.
A question-answering pipeline generally begins with question analysis. The system may identify the question type, important entities, constraints and expected answer form. Information Retrieval then identifies potentially relevant documents or passages, while a reader or reasoning component determines which information addresses the question. Some systems perform reranking so that the most promising passages receive additional processing. A final answer-generation component may extract an exact phrase, combine information from several sources or generate a natural-language response.
More advanced architectures may invoke databases, calculators, search tools or external APIs when answering requires information or computation not directly available in the text. In multi-step questions, the system may need to decompose the question into subquestions, retrieve evidence for each part and combine the results while maintaining provenance and logical consistency.
Question answering is difficult because natural-language questions can be ambiguous, underspecified or context-dependent. A question such as “When was it released?” may be impossible to answer without knowing what “it” refers to. Questions can also require temporal reasoning, numerical calculation, comparison, aggregation or inference across multiple documents. Some questions ask for facts explicitly stated in a source, while others require combining several facts to derive an answer.
Coreference resolution, Named-Entity Recognition, semantic parsing and relation extraction can therefore contribute to question answering. The system may also need to distinguish between an answer that is directly supported by evidence and one that is merely plausible. In high-quality systems, uncertainty handling and the ability to decline unsupported questions can be as important as the ability to produce answers.
Evaluation of Question Answering Systems depends strongly on the task. Exact Match measures whether a predicted answer matches an accepted reference exactly, while token-level F1 can provide partial credit when an answer overlaps with the reference. For generative systems, human evaluation may examine factual correctness, relevance, completeness, coherence and faithfulness to evidence. Retrieval components can be evaluated separately using recall and ranking metrics, because an answering model cannot select the correct answer from a passage that was never retrieved.
Multi-document systems require evaluation of evidence coverage and synthesis as well as final wording. A system can therefore produce a correct-looking answer for the wrong reasons or produce an incorrect answer despite retrieving relevant evidence. End-to-end evaluation should consequently be supplemented with component-level analysis and, where possible, provenance verification.
Modern Question Answering Systems frequently use large language models together with retrieval and tools. Retrieval-augmented generation can supply passages from a current document collection, allowing the model to answer questions about material that was not included in its original training data. Tool-augmented systems can query databases, execute calculations or access specialised services before generating an answer.
These architectures can reduce some forms of unsupported generation but introduce new dependencies: retrieval may return irrelevant passages, tools may fail, and the language model may misinterpret the retrieved evidence. Citation and provenance mechanisms can make answers easier to verify by identifying the documents or passages supporting individual claims. However, the presence of a citation does not automatically prove that the generated statement is correctly supported, so evidence alignment remains an important evaluation problem.
Question Answering Systems differ from Chatbots because question answering describes an information-access objective, whereas a chatbot is a broader conversational application that may answer questions as one of several functions. They differ from Information Retrieval because retrieval returns relevant resources while question answering attempts to provide a direct response derived from those resources or another knowledge source.
They differ from Natural Language Generation because generation describes the production of language, while question answering additionally requires identifying and satisfying the user's information need. Their main advantages are directness and reduced user effort, but limitations include hallucination, ambiguity, outdated knowledge, retrieval failures and difficulties with complex reasoning. Reliable systems therefore benefit from strong retrieval, explicit evidence handling, appropriate tool use, uncertainty management and evaluation against realistic questions rather than relying solely on linguistic fluency.
